IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO.

6, DECEMBER 2015 761

A Knee Point-Driven Evolutionary Algorithm

for Many-Objective Optimization
Xingyi Zhang, Ye Tian, and Yaochu Jin, Senior Member, IEEE

Abstract—Evolutionary algorithms (EAs) have shown to be a powerful framework for solving MOPs, since they can
be promising in solving many-objective optimization prob- find a set of Pareto optimal solutions in a single run. A large
lems (MaOPs), where the performance of these algorithms heavily number of multiobjective evolutionary algorithms (MOEAs)
depends on whether solutions that can accelerate convergence
toward the Pareto front and maintaining a high degree of diver- have been developed, e.g., NSGA-II [6], SPEA2 [7], IBEA [8],
sity will be selected from a set of nondominated solutions. In MOEA/D [9], PESA-II [10], and M-PAES [11], just to name a
this paper, we propose a knee point-driven EA to solve MaOPs. few. In all these MOEAs, a variety of selection strategies have
Our basic idea is that knee points are naturally most preferred been proposed to achieve fast convergence and high diversity,
among nondominated solutions if no explicit user preferences which play the most important role in determining the effec-
are given. A bias toward the knee points in the nondomi-
nated solutions in the current population is shown to be an tiveness and efficiency of the MOEA in obtaining the Pareto
approximation of a bias toward a large hypervolume, thereby optimal solutions.
enhancing the convergence performance in many-objective opti- Among various selection strategies, the Pareto-based non-
mization. In addition, as at most one solution will be identified dominated sorting approaches are the most popular, where
as a knee point inside the neighborhood of each solution in solutions having a better Pareto rank in the parent popula-
the nondominated front, no additional diversity maintenance
mechanisms need to be introduced in the proposed algorithm, tion or a combination of the parent and offspring populations
considerably reducing the computational complexity compared are selected. In addition to the dominance-based criterion, a
to many existing multiobjective EAs for many-objective opti- secondary criterion, often a diversity-related strategy, will be
mization. Experimental results on 16 test problems demonstrate adopted to achieve an even distribution of the Pareto optimal
the competitiveness of the proposed algorithm in terms of both solutions. NSGA-II [6] and SPEA2 [7] are two representa-
solution quality and computational efficiency.
tive Pareto-based MOEAs, which have been shown to be very
Index Terms—Convergence, diversity, evolutionary multiobjec- effective in solving MOPs having two or three objectives.
tive optimization, hypervolume (HV), knee point, many-objective However, the efficiency of such Pareto-based MOEAs will
optimization.
seriously degrade when the number of objectives is more than
I. I NTRODUCTION three, which are often known as many-objective optimization
problems (MaOPs).
ULTIOBJECTIVE optimization problems (MOPs) are
M commonly seen in real-world applications, especially
in the areas of engineering, biology, and economics [1]–[5].
MaOPs are widely seen in real-world applications
(see [12] and [13]). Increasing research attention has there-
fore been paid to tackling MaOPs in recent years, as it
Such optimization problems are characterized by multiple
has been shown that MaOPs cannot be solved efficiently
objectives that conflict with each other. Due to the con-
using MOEAs developed for solving MOPs with two or three
flicting nature of the objectives, usually no single optimal
objectives [14]–[17]. For example, NSGA-II performs very
solution exists; instead, a set of trade-off solutions, known as
well on MOPs with two or three objectives; however, its
Pareto optimal solutions can be found for MOPs. Over the
performance will dramatically deteriorate when the MOPs
past two decades, evolutionary algorithms (EAs) and other
have more than three objectives [18]. The main reason for
population-based meta-heuristics have been demonstrated to
this performance deterioration is that the selection criterion
Manuscript received June 15, 2014; revised September 30, 2014; accepted based on the standard dominance relationship fails to distin-
November 28, 2014. Date of publication December 4, 2014; date of current guish solutions in a population already in the early stage of
version November 25, 2015. This work was supported in part by the National
Natural Science Foundation of China under Grants 61272152, 61033003, the search, since most of the solutions in the population are
61202011, and 61472002; in part by the Natural Science Foundation of Anhui nondominated, although some of them may have a better abil-
Higher Education Institutions of China under Grant KJ2012A010; in part by ity to help the population to converge to the Pareto optimal
the Joint Research Fund for Overseas Chinese, Hong Kong; and in part by the
Macao Scholars of the National Natural Science Foundation of China under front [19]. Once the dominance-based selection criterion is not
Grant 61428302. able to distinguish solutions, MOEAs will often rely on a sec-
X. Zhang and Y. Tian are with Key Laboratory of Intelligent Computing ondary criterion, usually a metric for population diversity. As a
and Signal Processing of Ministry of Education, School of Computer Science
and Technology, Anhui University, Hefei 230039, China. result, MOEAs may end up with a set of well-distributed non-
Y. Jin is with the Department of Computing, University of Surrey, dominated solutions, which are unfortunately far from Pareto
Guildford, Surrey, GU2 7XH, U.K., and also with the College of Information optimal.
Sciences and Technology, Donghua University, Shanghai 201620, China
(e-mail: [email protected]). To enhance the ability of MOEAs to converge to the Pareto
Digital Object Identifier 10.1109/TEVC.2014.2378512 front, a variety of ideas have been proposed, which can be
1089-778X  c 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 762 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO. 6, DECEMBER 2015

largely divided into three categories [20], [21]. The first group in addition to the nondominance selection criterion. Therefore,
of ideas is to modify the traditional Pareto dominance defini- the proposed KnEA belongs to the second class of MOEAs
tion to increase the selection pressure toward the Pareto front. discussed above for solving MaOPs. Note, however, that by
This type of ideas has been widely adopted for solving MaOPs, knee points, we do not mean the knee points of the theoret-
such as -dominance [22], [23], L-optimality [24], fuzzy dom- ical (true) Pareto front; instead, we mean the knee points of
inance [25], and preference order ranking [26]. Compared with the nondominated fronts in the current population during the
MOEAs using the traditional Pareto dominance relationship, search process. Since at most one knee point is identified in
these strategies have been shown to considerably improve the the neighborhood of each solution, a preference over the knee
performance of MOEAs for solving MaOPs, although they are points also promotes diversity of the population, and conse-
very likely to converge into a sub-region of the Pareto front. quently no additional measures need to be taken in KnEA in
The second category of the ideas aims to combine the environmental selection. Note that calculating a diversity mea-
traditional Pareto dominance-based criterion with additional sure such as the crowding distance in NSGA-II can be highly
convergence-related metrics. Based on these ideas, solutions time-consuming for MaOPs.
are selected first based on the dominance relationship, and New contributions of this paper can be summarized as
then on the convergence-related metric. For example, some follows.
substitute distances based on the degree to which a solution 1) A knee point-driven MOEA has been suggested, where
is nearly dominated by any other solutions were proposed knee points of the nondominated fronts in the current
by Köppen and Yoshida [27] to improve the performance population are preferred in selection. We show that pref-
of NSGA-II. In [28], a binary -indicator based preference erence over knee points can approximately be seen as a
is combined with dominance to speed up convergence of bias toward larger HV, which is therefore very effective
NSGA-II for solving MaOPs. A grid dominance-based met- in both accelerating the convergence of the population
ric was also defined by Yang et al. [29] based on which an to the Pareto optimal front and maintaining diversity
effective MOEA, termed GrEA, for MaOPs has been proposed. of the solutions. We should stress that a large body of
The third type of ideas is to develop new selection cri- research work has been performed on identifying knee
teria based on some performance indicators. Three widely points in solving MOPs, most of which, however, con-
used performance indicator based MOEAs are IBEA [8], centrated on how to accurately find the knee points or
SMS-EMOA [30], and HypE [31]. IBEA uses a prede- local knee regions of the true Pareto front. To the best
fined optimization goal to measure the contribution of each of our knowledge, no work has been reported on using
solution, while SMS-EMOA and HypE are based on the knee points as the secondary criterion to enhance the
hypervolume (HV) value. search performance of MOEAs for MaOPs.
There are also a large number of other many-objective opti- 2) Within the KnEA, an adaptive strategy for identifying
mization algorithms, which adopt different ideas from those knee points in a small neighborhood, i.e., local knee
discussed above. For example, some researchers attempted to regions, has been proposed without using prior knowl-
solve MaOPs by using a reduced set of objectives [32], [33], edge about the number of knee points in the true Pareto
while others suggested to use interactive user preferences [34] front. The purpose of the adaptive strategy is not to
or reference points [35] during the search. An interesting find precisely the knee points of the true Pareto front;
MOEA for solving MaOPs, called NSGA-III, was also based instead, it is meant to locate those local knee points
on a set of reference points [36], where nondominated solu- of the nondominated fronts in the population combin-
tions close to the reference points are prioritized. Note also ing the parent and offspring populations at the present
that some MOEAs have shown to perform fairly well for some generation to accelerate the convergence and promote
MaOP test problems [37], such as the decomposition-based diversity.
multiobjective evolutionary algorithm, termed MOEA/D [9], 3) Extensive experimental results have been conducted to
although they are not specifically designed for solving verify the performance of KnEA for solving MaOPs by
MaOPs. comparing it with several state-of-the-art MOEAs for
In multiobjective optimization, knee points are a sub-set of MaOPs on two suites of widely used test problems.
Pareto optimal solutions for which an improvement in one Our results demonstrate that KnEA outperforms the
objective will result in a severe degradation in at least another compared MOEAs for MaOPs in terms of two widely
one. For MOPs, solutions in the knee region of the Pareto front used performance indicators. Moreover, KnEA is com-
will be naturally preferred if no other user-specific or problem- putationally much more efficient than two of the three
specific preferences are available. As previously discussed, compared Pareto-based MOEAs and comparable to the
most existing MOEAs do not work efficiently for MaOPs rest one, although it is slightly inferior to MOEA/D,
mainly due to the loss of selection pressure because most or which is known for its high computational efficiency
all solutions in the population are nondominated already in a for MaOPs.
very early search stage. In this paper, we propose a knee point- The rest of this paper is organized as follows. In Section II,
driven evolutionary algorithm (KnEA), in which preferences existing work related to the identification of knee points in
are given to knee points among the nondominated solutions multiobjective optimization is discussed and the motivation
in selection. In other words, knee points are used as the sec- of using knee points as a selection criterion is justified. The
ondary criterion for selecting parents for the next generation details of the proposed KnEA for MaOPs are described in

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: KNEE POINT-DRIVEN EA FOR MANY-OBJECTIVE OPTIMIZATION 763

Section III. Simulation results are presented in Section IV to

empirically compare search performance and runtime of the
KnEA with four state-of-the-art methods for MaOPs. Finally,
the conclusion is given in Section V.

II. R ELATED W ORK AND M OTIVATION

In this section, we first review the related work on finding
knee points in evolutionary multiobjective optimization. Then,
we elaborate the motivation of using knee points detected dur-
ing the search for driving the population toward the Pareto
optimal front and maintaining population diversity.

A. Related Work Fig. 1. Illustration of the motivation of KnEA. B can be seen as a knee
A large number of MOEAs have been proposed to find point among the five nondominated solutions A, B , B, C, and D. Selecting
solution B, the knee point can be more beneficial than B in terms of HV.
local regions or points of interest in the Pareto optimal
front [38]–[40]. Among various preferences, knee points are
often considered to be of interest in the Pareto optimal front knee points were used as mobile reference points and the
and much research work has been dedicated to finding knee search of the algorithm was guided toward these points.
points or knee regions (neighboring regions of knee points) KR-NSGA-II has been shown to perform well in controlling
using MOEAs. the extent of knee regions for MOPs with two or more objec-
Intuitively, a knee point in the Pareto optimal front refers tives, assuming that some prior information on the number of
to the solution with the maximum marginal rates of return, knee points of the MOP is known. Deb and Gupta [48] sug-
which means that a small improvement in one objective of gested several new definitions for identifying knee points and
such a solution is accompanied by a severe degradation in at knee regions for bi-objective optimization problems. The pos-
least another. As knee points are naturally preferred, several sibility of applying such methods to solve bi-objective engi-
multiobjective optimization algorithms have been developed to neering problems has also been discussed. Branke et al. [42]
find the knee points or knee regions in the Pareto optimal front presented two test problems with knee points, one bi-objective
instead of approximating the whole front. Das [41] suggested and one three-objective, to evaluate the ability of MOEAs for
a method to find the knee points in the Pareto front based finding knee points. Tušar and Filipič [49] extended these two
on normal boundary intersection, which has been shown to be test problems to four-objective and five-objective optimization
very efficient for characterizing knee points. Branke et al. [42] problems.
proposed two variants of NSGA-II for finding knee regions, Various definitions for characterizing knee points and knee
where the crowding distance in NSGA-II was substituted by regions have been suggested (see [41]–[44], [48]). In this
two new measures: 1) an angle-based measure and 2) a utility paper, we adopt the definition presented by Das [41] and
measure. The variant with the utility measure can be used for Bechikh et al. [46], which will be further discussed in
problems with any number of objectives, while the one with Section III-C.
the angle-based measure works only for bi-objective problems.
These two variants of NSGA-II have been shown to perform B. Motivation of This Paper
very well on finding knee regions in the Pareto front, which, As can be seen in the discussions above, the importance
however, are not able to control the spread of these regions. of knee points and knee regions has long been recognized
To control the spread of knee regions, in evolutionary multiobjective optimization. Nevertheless, the
Rachmawati and Srinivasan [43], [44] developed an MOEA use of knee points to improve the search ability of MOEAs,
based on a weighted sum niching method, where the extent especially for solving MaOPs, has not been reported so far. In
and the density of convergence of the knee regions were this paper, we hypothesize that the search ability of MOEAs
controlled by the niche strength and the total number of for solving MaOPs can be significantly enhanced by giving
solutions in the region, known as pool size. However, such preferences to the knee points among nondominated solutions.
control on the extent and the density of the knee regions is To elaborate this hypothesis, consider five nondominated
very rough. Schütze et al. [45] presented two strategies for solutions of a bi-objective optimization problem, A(1, 16),
finding knee points and knee regions that can be integrated B (6, 11), B(7, 7), C(11, 6), and D(16, 1), where the two ele-
into any stochastic search algorithms. Experimental results ments of a solution indicate the values of the two objectives,
illustrated that these two strategies were very efficient in as shown in Fig. 1. From Fig. 1, we can see that solution B
finding the knee points and knee regions of bi-objective can be considered as a knee point of the nondominated front
optimization problems. However, these methods are not consisting of the five nondominated solutions. Assume that
extendable to MOPs with more than two objectives. four solutions are to be selected from the five nondominated
Bechikh et al. [46] proposed an MOEA for finding knee solutions for next population. Since these five solutions are all
regions, termed KR-NSGA-II by extending the reference point nondominated, a secondary criterion must be used for select-
based MOEA, called R-NSGA-II [47]. In KR-NSGA-II, the ing four out of the five solutions. If a diversity-based criterion,

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 764 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO. 6, DECEMBER 2015

TABLE I
R ELATIONSHIP B ETWEEN D ISTANCE OF B TO THE E XTREME L INE AD
Algorithm 1 General Framework of KnEA
W ITH THE HV OF S OLUTION S ET C ONSISTING OF A, B, C, AND D Require: P (population), N (population size), K (set of knee
points), T (rate of knee points in population)
1: r ← 1, t ← 0 /*adaptive parameters for finding knee
points*/
2: K ← ∅
3: P ← Initialize(N)
4: while termination criterion not fulfilled do
for example, the crowding distance defined in [6] is used for 5: P ← Mating_selection(P,
 K, N)
selection, then solutions A, B , C, and D will be selected. If we 6: P ← P Variation(P , N)
replace solution B with the knee point B, the selected solution 7: F ← Nondominated_sort(P) /*find the solutions in the
set will be A, B, C, and D. first i fronts Fj , 1 ≤ j ≤ i, where i is the minimal value
Let us now compare the quality of the above two solution such that |F1 ∪ . . . ∪ Fi | >= N */
sets using the HV, which is one of the most widely used per- 8: [K, r, t] ← Finding_ knee_ point(F, T, r, t)
formance indicators in multiobjective optimization [50]. For 9: P ← Environmental_selection(F, K, N)
calculating the HV of the two sets, we assume that the ref- 10: end while
erence point is (18, 18). In this case, the HV of the solution 11: return P
set consisting of A, B , C, and D is 139, while the HV of the
solution set consisting of A, B, C, and D is 150.
From the above illustrative example, we can observe that a good diversity. By contrast, diversity is explicitly embedded
selecting knee points can be more beneficial than selecting in the knee point identification process proposed in this paper,
more diverse solutions in terms of the HV. To take a closer since at most one solution will be labeled as a knee point
look at the relationship between the position of point B and the in the neighborhood of a solution. The above hypothesis has
HV of the solution set consisting of A, B, C, and D, we move been verified by our empirical results comparing the proposed
the position of B from B(6, 7), which is the leftmost possi- method with HypE, a HV-based method. Refer to Section IV
ble position, to B(11, 7), which is rightmost possible position for more details.
to maintain the nondominated relationships between B and
the other four solutions. Now, we examine the relationship
between the distance of B to the extreme line AD, which is III. P ROPOSED A LGORITHM FOR M ANY-O BJECTIVE
described by f1 + f2 = 17, and the HV of the solution set O PTIMIZATION
consisting of A, B, C, and D on five different positions. The KnEA is in principle an elitist Pareto-based MOEA.
results are listed in Table I. The main difference between KnEA and other Pareto-based
From Table I, we can see that when B moves from B(6, 7) to MOEAs such as NSGA-II is that knee points are used as a
B(7, 7), the HV of the solution set consisting of A, B, C, and D secondary selection criterion in addition to the dominance rela-
decreases from 159 to 150, while the distance to the extreme tionship. During the environmental selection, KnEA does not
line decreases from 2.83 to 2.12. When point B further moves use any explicit diversity measure to promote the diversity of
to the right to B(74/9, 7), the HV drops to 139, which is equal the selected solution set. In the following, we describe the
to the HV of the solution set consisting of A, B , C, and D. In main components of KnEA.
this case, the distance of point B to the extreme line is further
reduced to 1.26 and B is no longer a typical knee point. If
point B continues to move to B(10, 7), B is exactly located A. General Framework of the Proposed Algorithm
on the extreme line, and the HV of solution set consisting of The general framework of KnEA is similar to that of
A, B, C, and D becomes 123, which is even smaller than that NSGA-II [6], which consists of the following main compo-
of solution set consisting of A, B , C, and D. Therefore, we nents. First, an initial parent population of size N is randomly
can conclude that the more typical B is a knee point, the more generated. Second, a binary tournament strategy is applied to
likely it will contribute to a large HV. select individuals from the parent population to generate N
From the above example, we can hypothesize that a pref- offspring individuals using a variation method. In the binary
erence over knee points can be considered as an approxima- tournament selection, three tournament metrics are adopted,
tion of the preference over larger HVs. Compared with the namely, the dominance relationship, the knee point criterion,
HV-based selection, however, knee point-based selection offers and a weighted distance measure. Third, nondominated sorting
the following two important advantages. First, the identifica- is performed on the combination of the parent and offspring
tion of knee points is computationally much more efficient population, followed by an adaptive strategy to identify solu-
than calculating the HV, in particular when the number of tions located in the knee regions of each nondominated front in
objectives is large. To be more specific, the computational the combined population. Fourth, an environmental selection
time for calculating the HV increases exponentially as the is conducted to select N individuals as the parent population
number of objectives increases, while the time for identify- of the next generation. This procedure repeats until a termina-
ing knee points increases only linearly. Second, although the tion condition is met. The above main components of KnEA
HV implicitly takes diversity into account, it cannot guarantee are presented in Algorithm 1.

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: KNEE POINT-DRIVEN EA FOR MANY-OBJECTIVE OPTIMIZATION 765

Algorithm 2 Mating_Selection(P, K, N)
Require: P (population), K (set of knee points), N (popu-
lation size)
1: Q ← ∅
2: while |Q| < N do
3: randomly choose a and b from P
4: if a ≺ b then
5: Q ← Q {a}
6: else if b ≺ a then
7: Q ← Q {b}
8: else
9: if a ∈ K andb∈ / K then
10: Q ← Q {a}
11: else if a ∈/K and b ∈ K then
12: Q ← Q {b}
Fig. 2. Illustrative example where the proposed weighted distance may be
13: else advantageous over the crowding distance. In this example, neither solution
14: if DW(a) > DW(b) then B nor C will have the chance to win against other solutions if the crowding
15: Q ← Q {a} distance is adopted. Both B and C have a chance to win according to the
defined weighted distance.
16:
 < DW(b) then
else if DW(a)
17: Q ← Q {b} an equal weighted distance, then one of them will be randomly
18: else chosen for reproduction.
19: if rand() <0.5 then A weighted distance is designed for choosing a winning
20: Q ← Q {a} solution in the binary tournament mating selection if neither
21: else  the dominance comparison nor the knee point criterion can
22: Q ← Q {b} distinguish the two solutions involved in the tournament. We
23: end if adopt here the weighted distance measure to address some
24: end if potential weakness of the crowding distance metric proposed
25: end if in NSGA-II [6]. Fig. 2 illustrates a situation, where if the
26: end if crowding distance is used, neither solution B nor solution C
27: end while
will have the chance to win against other solutions. However,
28: return Q
from the diversity point of view, it would be helpful if either
B or C can have a chance to win in the tournament for
reproduction.
In the following, we describe in detail the binary tournament The weighted distance of a solution p in a population based
mating selection, the adaptive knee point detection method, on the k-nearest neighbors is defined as
and the environmental selection, which are three important 
k
components in KnEA. DW(p) = wpi disppi (1)
i=1
rp
wpi = k i (2)
B. Binary Tournament Mating Selection i=1 rpi
The mating selection in KnEA is a binary tournament 1
rpi =  k  (3)
selection strategy using three tournament strategies, namely,  
disppi − 1
k i=1 disppi 
dominance comparison, knee point criterion, and a weighted
distance. Algorithm 2 describes the detailed procedure of the where pi represents the ith nearest neighbor of p in the
mating selection strategy in KnEA. population, wpi represents the weight of pi , disppi represents
In the binary tournament mating selection in KnEA, two the Euclidean distance between p and pi , and rpi represents
individuals are randomly chosen from the parent population. the rank of distance disppi among all the distances disppj ,
If one solution dominates the other solution, then the former 1 ≤ j ≤ k. From (3), it can be seen that a neighbor of
solution is chosen, referring to lines 4–7 in Algorithm 2. If p will have a larger rank if it is nearer to the center of
the two solutions are nondominated with each other, then the all considered neighbors of p. By using the above weighted
algorithm will check whether they are both knee points. If only distance, we can verify that both solutions B and C have
one of them is a knee point, then the knee point is chosen, see- a certain probability to be selected in tournament selection.
ing lines 9–12 in Algorithm 2. If both of them are knee points Note that some existing distance metrics can also address the
or neither of them is a knee point, then a weighted distance above weakness of the crowding distance, such as the grid
will be used for comparing the two solutions, as described crowding distance (GCD) proposed in GrEA [29]. Compared
in lines 14–17 in Algorithm 2. The solution with the larger to GCD, the weighted distance presented above is easier to
weighted distance wins the tournament. If both solutions have calculate.

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 766 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO. 6, DECEMBER 2015

For minimization problems, only solutions in the convex

knee regions are of interest. Therefore, the distance measure
in (4) can be modified as follows to identify knee points:
⎧ |ax +by +c|
⎨ √ A A
if axA + byA + c < 0
a2 +b2
d(A, L) = (5)
⎩ − |ax√A +byA +c|
otherwise.
2 2 a +b
The above distance measure for identifying knee points
can be easily extended to optimization problems with more
than two objectives, where the extreme line will become a
hyperplane.
The example in Fig. 3 indicates that the size of neighbor-
hood of the solutions will heavily influence the results of the
identified knee points. Given the size of the neighborhood
defined in Fig. 3, solutions B, E, and G are identified as knee
Fig. 3. Illustration for determining knee points in KnEA for a bi-objective
minimization problem. In this example, solutions B, E, and G are identified as points. Imagine, however, that if all solutions are included in
knee points for the given neighborhood denoted by the rectangles in dashed the same neighborhood of a solution, then only solution E will
lines. be identified as knee point. For this reason, a strategy to tune
the size of the neighborhood of solutions is proposed, which
will be described in the following.
C. Adaptive Strategy for Identifying Knee Points
Assume the combined population at generation g con-
Knee points play a central role in KnEA. The knee points tains NF nondominated fronts, each of which has a set of
are used as a criterion only next to the dominance criterion in nondominated solutions denoted by Fi , 1 ≤ i ≤ NF . The
both mating and environmental selection. Therefore, an effec- neighborhood of a solution is defined by a hyper cube of size
tive strategy for identifying solutions in the knee regions of j
R1g × R2g × · · · × Rg × · · · × RM g , where 1 ≤ j ≤ M,
the nondominated fronts in the combined population is crit- M is the number of objectives. Specifically, the size of the
ical for the performance of KnEA. To this end, an adaptive j
neighborhood regarding the jth objective, Rg , is determined as
strategy is proposed for finding knee points in the population
combining the parent and offspring populations at the present Rjg = f maxjg − f mingj · rg (6)
generation.
j j
Fig. 3 presents an example for illustrating the main idea for where f maxg and f ming denote the maximal and the minimal
determining knee points in the proposed strategy, where the values of the jth objective at the gth generation in set Fi , and
nondominated front of a bi-objective minimization problem in rg is the ratio of the size of the neighborhood to the span of the
consideration consists of nine solutions. First of all, an extreme jth objective in nondominated front Fi at the gth generation,
line L is defined by the two extreme solutions, one having the which is updated as
maximum of f1 and the other having the maximum of f2 among 1−tg−1 /T
all the solutions in the nondominated front. Then, we calculate rg = rg−1 ∗ e− M (7)
the distance of each solution to L. A solution is identified as a where rg−1 is the ratio of the size of the neighborhood to the
knee point if its distance to the extreme line is the maximum span of the jth objective of the solutions in Fi at the (g − 1)th
in its neighborhood. generation, M is the number of objectives, tg−1 is the ratio of
By looking at Fig. 3, we can see that solution B is knee points to the number of nondominated solutions in front
a knee point in its neighborhood denoted by the rectan- Fi at the (g − 1)th generation, and 0 < T < 1 is a threshold
gle in dashed lines, as it has the maximum distance to L that controls the ratio of knee points in the solution set Fi .
among A, B, C, and D inside its neighborhood. Intuitively, Equation (7) ensures that rg will significantly decrease when
solution E is also a knee point compared with solution F in tg−1 is much smaller than the specified threshold T, and the
its neighborhood. Note that, if there is only one solution in decrease of rg will become slower as the value of tg−1 becomes
its neighborhood, e.g., solution G in Fig. 3, this solution will larger. rg will remain unchanged when tg−1 reaches the given
also be considered as a knee point. The above knee point def- threshold T. tg and rg are initialized to 0 and 1, respectively,
inition leads to the benefit that the diversity of the population i.e., t0 = 0 and r0 = 1.
is implicitly taken into account. Fig. 4 presents the change of parameters R1g , rg , and tg
The use of distance to the extreme line L to characterize on DTLZ2 with three objectives as the evolution proceeds,
knee points was first proposed by Das [41]. For a bi-objective where T is set to T = 0.5. The size of the neighborhoods is
minimization problem, L can be defined by ax + by + c = 0, adapted according to the ratio of the identified knee points to
where the parameters can be determined by the two extreme the total number of nondominated solutions. In the early stage
solutions. Then, the distance from a solution A(xA , yA ) to L of the evolutionary optimization, the size of neighborhoods
can be calculated as will decrease quickly, and thus the number of found knee
|axA + byA + c| points will significantly increase. The ratio of knee points to
d(A, L) = √ . (4)
a2 + b2 all nondominated solutions (tg ) will increase as the evolution

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: KNEE POINT-DRIVEN EA FOR MANY-OBJECTIVE OPTIMIZATION 767

Algorithm 4 Environmental_Selection(F, K, N)
Require: F (sorted population), K (set of knee points),
N (population size)
1: Q ← ∅ /*next population*/
 
2: Q ← F1
 . . . Fi−1
3: Q ← Q (K Fi )
4: if |Q| > N then
5: delete |Q|−N solutions from Q which belong to K Fi
and have the minimum distances to the hyperplane
6: else if |Q| < N then
7: add N − |Q| solutions from Fi \(K Fi ) to Q which
have the maximum distances to the hyperplane
8: end if
Fig. 4. Example of the changes of the parameters R1g , rg , and tg of the first
9: return Q
front over the number of generations on DTLZ2 with three objectives.

Algorithm 3 Finding_Knee_Point(F, T, r, t)
Require: F (sorted population), T (rate of knee points in proposed adaptive strategy aims to find out the knee solutions
population), r, t (adaptive parameters) in the neighborhoods, which will be preferred in the mating
1: K ← ∅ /* knee points */ and environmental selection. Note again that by knee points
2: for all Fi ∈ F do here, we do not mean the knee points of the true Pareto front;
3: E ← Find_extreme_solution(Fi ) /* Fi denotes the set instead, we mean the knee points of the nondominated fronts in
of solutions in the ith front */ the combined population at the current generation. In addition,
4: L ← Calculate_extreme_hyperplane(E) some of the solutions identified as knee points may not be true
5: update r by formula (7) knee points, which, however, can speed up the convergence
6: f max ← maximum value of each objective in Fi performance and enhance the diversity of the population.
7: f min ← minimum value of each objective in Fi
8: calculate R by formula (6)
9: calculate the distance between each solution in Fi and D. Environmental Selection
L by formula (5) Environmental selection is to select fitter solutions as par-
10: sort Fi in a descending order according to the distances ents for the next generation. Similar to NSGA-II, KnEA selects
11: SizeFi ← |Fi | parents for the next generation from a combination of the
12: for all p ∈ Fi do parent and offspring populations of this generation, which
j j
13: NB ← {a|a
 ∈ Fi → | fa − fp | ≤ R , 1 ≤ j ≤ M}
j
therefore is an elitist approach. Whereas both NSGA-II and
14: K ← K {p} KnEA adopt the Pareto dominance as the primary criterion in
15: Fi ← Fi \NB environmental selection, KnEA prefers knee points instead of
16: end for the nondominated solutions with a larger crowding distance
17: t = |K|/SizeFi as NSGA-II does. Algorithm 4 presents the main steps of
18: end for environmental selection in KnEA.
19: return K, r and t Before environmental selection, KnEA performs nondomi-
nated sorting using the efficient nondominated sorting (ENS)
algorithm reported in [51], forming NF nondominated fronts,
proceeds, which, in the meantime, will gradually decrease the Fi , 1 ≤ i ≤ NF . Similar to NSGA-II, KnEA starts to select the
size of the neighborhoods. When tg is close to the threshold T, nondominated solutions in the first nondominated front (F1 ).
the size of the neighborhoods will remain constant. If the number of solutions in F1 is larger than the population
The main steps of the adaptive strategy for detecting knee size N, which is very likely already in the early generations
points are presented in Algorithm 3. The same procedure can in many-objective optimization, then knee points in F1 are
be repeated for all nondominated fronts in the combined pop- selected first as parents for the next population. Let the num-
ulation until knee points are identified for all nondominated ber of knee points in F1 be NP1 . In case NP1 is larger than N,
fronts. Note, however, that in the late search stage of MOPs, then N knee points having a larger distance to the hyperplane
and actually already in the early stages of MaOPs, we only are selected, referring to line 5 in Algorithm 4. Otherwise,
need to find the knee points in the first front due to the large NP1 knee points are selected together with (N − NP1 ) other
number of nondominated solutions present in this front. solutions in F1 that have a larger distance to the hyperplane
From the above descriptions, we can find that the proposed of F1 .
adaptive knee point identification algorithm differs consid- If the number of solutions in F1 is smaller than N,
erably from the existing methods for finding knee points. KnEA turns to the second nondominated front for selecting
Whereas, most existing MOEAs for finding the knee points are the remaining (N − |F1 |) parent solutions. If |F2 | is larger
to accurately locate the knee points in the true Pareto front, the than N − |F1 |, then the same procedure described above

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 768 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO. 6, DECEMBER 2015

and polynomial mutation [53], are performed on each deci-

sion variable of the parent solutions, which needs a runtime
of O(DN) to generate N offspring, where D is the number of
decision variables. Nondominated sorting needs a runtime of
O(MN 2 ) in the worst case for the combined population of size
2N for optimization problems with M objectives. Knee point
identification consists of the following two operations. First,
obtaining the hyperplane and calculating the distance between
each nondominated solution and the hyperplane, which at most
needs a runtime of O(MN). Second, checking whether the non-
dominated solutions are knee points in their neighborhoods,
which costs a runtime of O(MN 2 ). Therefore, knee point iden-
tification takes at most a runtime of O(MN 2 ) in total. For
environmental selection, a runtime of O(NlogN) is needed,
Fig. 5. Example showing the number of solutions in the first nondom- since the most time-consuming step is to sort the nondomi-
inated front, together with the number of solutions selected based on the
knee point criterion and the distance to the hyperplane criterion. The results nated solutions according to their distances to the hyperplane.
are obtained on the three-objective DTLZ2 using a population size of 100, Therefore, KnEA needs at most a total runtime of O(GMN 2 ),
i.e., the combined population size is 200. where G is the number of generations.
Compared with most popular MOEAs for MaOPs, KnEA is
computationally very efficient. A theoretical comparison of the
will applied to F2 . This process is repeated until the parent computational time of KnEA with these algorithms is beyond
population for the next generation is filled up. the scope of this paper; however, we will empirically compare
It would be of interest to know how many solutions in the the runtime performance of KnEA with four state-of-the-art
combined population are nondominated, how many are iden- MOEAs for MaOPs, details of which will be presented in the
tified as knee points, and how many will be selected based next section.
on the distance to the hyperplane as the evolution proceeds.
Take the three-objective DTLZ2 as an illustrative example and
assume the population size is 100 and T is set to T = 0.5. IV. E XPERIMENTAL R ESULTS AND D ISCUSSION
Fig. 5 presents the number of solutions in the first non- In this section, we verify the performance of KnEA by
dominated front in the combined population at generations empirically comparing it with four popular MOEAs for
5, 10, 50, 100, and 250, where the number of identified knee MaOPs, namely, GrEA [29], HypE [31], MOEA/D [9], and
points and the number of solutions selected based on the NSGA-III [36]. The experiments are conducted on 16 test
distance to the hyperplane are indicated in black and gray, problems taken from two widely used test suites, DTLZ [54]
respectively. From the figure, we can see that the number of and WFG [55]. For each test problem, 2, 4, 6, 8, and 10 objec-
nondominated solutions is slightly less than 100 at genera- tives will be considered, respectively. We compare both the
tion 5 and thus all solutions in the first nondominated front quality of the obtained nondominated solution sets in terms of
will be selected. At generation 50, by contrast, almost all widely used performance indicators and the computational effi-
solutions (180 out of 200) are nondominated and a major- ciency with respect to runtime. Note that the ENS-SS reported
ity of the selected solutions (91 out of 100) are knee points. in [51] has been adopted as the nondominated sorting approach
We can imagine that as the number of objectives increases, in all compared MOEAs.
most selected solutions will be knee points even in early gen-
erations. These results indicate that the proposed method is A. Experimental Setting
different from the nondominance based selection and the dis- For a fair comparison, we adopt the recommended param-
tance based selection, and therefore the identified knee points eter values for the compared algorithms that have achieved
play an essential role in determining the performance of the the best performance. Specifically, the parameter setting for
algorithm. all conducted experiments are as follows.
1) Crossover and Mutation: The SBX [52] and polyno-
E. Empirical Computational Complexity Analysis mial mutation [53] have been adopted to create offspring. The
In this section, we provide an upper bound of the runtime distribution index of crossover is set to nc = 20 and the distri-
of KnEA. Within one generation, KnEA mainly performs the bution index of mutation is set to nm = 20, as recommended
following five operations: 1) mating selection; 2) genetic vari- in [56]. The crossover probability pc = 1.0 and the mutation
ations; 3) nondominated sorting; 4) knee-point identification; probability pm = 1/D are used, where D denotes the number
and 5) environmental selection. For a population size N and an of decision variables.
optimization problem of M objectives, mating selection needs 2) Population Sizing: To avoid that the generated reference
a runtime of O(MN 2 ) to form a mating pool of size N, as the points are all located along the boundary of Pareto fronts for
calculation of the weighted distances involves calculating the problems with a large number of objectives, the strategy of
distance between pairs of solutions in the population. Genetic two-layered reference points recommended in NSGA-III [36]
variations, here the simulated binary crossover (SBX) [52] was adopted to generate uniformly distributed weight vectors

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: KNEE POINT-DRIVEN EA FOR MANY-OBJECTIVE OPTIMIZATION 769

TABLE II TABLE III

S ETTING OF P OPULATION S IZE IN NSGA-III AND MOEA/D, W HERE PARAMETER S ETTING OF div IN G R EA ON
p1 AND p2 ARE PARAMETERS C ONTROLLING THE N UMBERS OF DTLZ AND WFG T EST S UITES
R EFERENCE P OINTS A LONG THE B OUNDARY OF THE
PARETO F RONT AND I NSIDE IT, R ESPECTIVELY

in NSGA-III and MOEA/D. Table II presents the setting of TABLE IV

PARAMETER S ETTING OF T IN K N EA ON
population size in NSGA-III and MOEA/D, where p1 and p2 DTLZ AND WFG T EST S UITES
are parameters controlling the numbers of reference points
along the boundary of the Pareto front and inside it, respec-
tively. For each test problem, the population size of HypE,
GrEA, and KnEA is set to the same as that of NSGA-III and
MOEA/D.
3) Number of Runs and Stopping Condition: We perform
20 independent runs for each algorithm on each test instance
on a PC with a 3.16 GHz Intel Core 2 Duo CPU E8500 and the
Windows 7 SP1 64 bit operating system. The number of itera-
tions is adopted as the termination criterion for all considered
TABLE V
algorithms. For DTLZ1 and WFG2, the maximum number of S ETTING OF T EST P ROBLEMS DTLZ1 TO DTLZ7
iterations is set to 700, and to 1000 for DTLZ3 and WFG1.
For DTLZ2, DTLZ4, DTLZ5, DTLZ6, DTLZ7, and WFG 3
to WFG9, we set the maximum number of iterations to 250.
4) Other Parameters: The parameter setting for div in
GrEA is taken from [29], which stands for the number of
divisions in each dimension in GrEA. The method for calcu-
lating HV suggested in [50] is adopted in HypE: the exact
method suggested in [50] is used to calculate the indicator
value for test instances with two objectives, and otherwise the
Monte Carlo sampling described in [31] is adopted to approx-
imately calculate the indicator, where 10 000 samples are used
in our experiments. For MOEA/D, the range of neighborhood
is set to N/10 for all test problems, and Tchebycheff approach
is employed as the aggregation function, where N is the pop- of the achieved nondominated solutions. However, as we dis-
ulation size. Three-nearest neighbors are used for calculating cussed in Section II-B, HV may have a bias toward typical
the weighted distance in KnEA, unless otherwise specified. knee points. Note that the larger the HV value is, the better
Table III lists the parameter setting of div in GrEA on DTLZ the performance of the algorithm. By contrast, a smaller IGD
and WFG test suites. To get the optimal setting for div, we value indicates better performance of the MOEA. In this paper,
tested many values for div for each of the test instances based (1, 1, . . . , 1) is chosen as the reference point in HV calcula-
on the recommendation in [29] and chose the one that pro- tion. For the objective values of WFG test problems to have
duced the best performance for GrEA. Table IV lists the setting the same scale, each of the objective values has been normal-
of T in KnEA on DTLZ and WFG test suites. As shown in ized between the interval [0, 1] before calculating the HV. In
the table, for DTLZ2, DTLZ4, and all test problems in the addition, since the exact calculation of HV is computation-
WFG suite except for WFG4 and WFG9, T is set to 0.6 for ally extremely intensive for MaOPs, the Monte Carlo method
problems with two objectives and 0.5 otherwise. is adopted for estimating the HV when the test problem has
5) Quality Metrics: Two widely used performance more than four objectives, where 1 000 000 sampling points
indicators, the HV [50] and the inverted generational are used. We should stress that most recently a computation-
distance (IGD) [57], [58] are used to evaluate the perfor- ally efficient method was also suggested to calculate the exact
mance of the compared algorithms. It is believed that these two HV [59]. On the other hand, IGD requires a reference set of
performance indicators can not only account for convergence Pareto optimal solutions, which are uniformly chosen from the
(closeness to the true Pareto front), but also the distribution true Pareto fronts of test problems. However, for different test

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 770 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO. 6, DECEMBER 2015

TABLE VI
IGD R ESULTS OF THE F IVE C OMPARED A LGORITHMS ON DTLZ1 TO DTLZ7, W HERE THE B EST M EAN FOR E ACH T EST
I NSTANCE IS S HOWN W ITH A G RAY BACKGROUND

problems with different numbers of objectives, it is impossi- The results on the seven DTLZ test problems are given in
ble to use exactly the same number of reference points. In this Table VI, with both the mean and standard deviation of the
paper, we set the number of reference points to the integer that IGD values averaged over 20 independent runs being listed for
is closest to 500. the five compared MOEAs, where the best mean among the
five compared algorithms is highlighted. From the table, we
B. Results on the DTLZ Suite can find that both MOEA/D, HypE and NSGA-III performed
The DTLZ test suite [54] is a class of widely used bench- well on DTLZ test problems with two objectives. Among the
mark problems for testing the performance of MOEAs. Seven seven DTLZ test problems, MOEA/D achieved the smallest
test functions, from DTLZ1 to DTLZ7 are used in the exper- IGD values on five bi-objective test problems, while HypE
iments here and their parameters are set as suggested in [54], and NSGA-III achieved an IGD value very close to the small-
which are presented in Table V. est one on all bi-objective DTLZ test problems. Note that

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: KNEE POINT-DRIVEN EA FOR MANY-OBJECTIVE OPTIMIZATION 771

MOEA/D obtained a worse IGD value on the bi-objective

DTLZ4. It appeared that MOEA/D does not work well on
DTLZ4 with any number of objectives. The main reason is
that DTLZ4 is a nonuniform MOP, which means that a set of
evenly distributed weight combinations will lead to nonuni-
formly distributed Pareto optimal solutions. This is a known
weakness of weighted aggregation methods for nonuniform
MOPs.
For DTLZ test problems with more than three objectives,
GrEA and NSGA-III performed better than MOEA/D and
HypE on all test problems except for DTLZ5 and DTLZ6.
MOEA/D and HypE worked very well on DTLZ5 and DTLZ6
with more than three objectives. HypE obtained the small-
est IGD value among the five MOEAs under comparison Fig. 6. Runtime(s) of the five algorithms on all DTLZ test problems, where
the runtime of an algorithm on M objectives is obtained by averaging over
on all DTLZ5 test instances with more than four objectives the runtimes consumed by the algorithm for one run on all DTLZ problems
and DTLZ6 with ten objectives, while MOEA/D obtained the with M objectives.
smallest IGD value on DTLZ6 with six and eight objectives
and obtained the second smallest IGD value on the remaining
test instances of DTLZ5 and DTLZ6 with more than three consumed by the algorithm for one run on all M-objective
objectives except for DTLZ5 with four objectives (on this DTLZ problems. Note that the runtimes are displayed in
instance, MOEA/D achieved a value very close to the second logarithm in the figure. As shown in the figure, MOEA/D
smallest IGD value). These empirical results may illustrate that outperforms the four compared MOEAs on all instances in
HypE and MOEA/D are well suited for dealing with MaOPs terms of runtime, which are much less than HypE, GrEA,
whose Pareto front is a degenerated curve. NSGA-III, and KnEA. Note, however, that although KnEA
Similar to GrEA and NSGA-III, the performance of KnEA consumed more time than MOEA/D did, it used much less
is also very promising on the seven DTLZ test problems time than GrEA and HypE and consumed comparable run-
with more than three objectives. For DTLZ2, DTLZ4, and time with NSGA-III. We see that KnEA took roughly only
DTLZ7 with more than three objectives, KnEA achieved a one-third of the runtime of GrEA on bi-objective instances.
slightly better IGD value than GrEA and NSGA-III on all test As the number of objectives increases, the runtime of KnEA
instances except for DTLZ2 with four objectives. For DTLZ5 increased only very slightly. For ten-objective test problems,
and DTLZ6 with more than three objectives, KnEA achieved the runtime of KnEA is only about one seventh of that of
a similar IGD value as GrEA and NSGA-III on DTLZ5, but it GrEA. Among the five algorithms under comparison, HypE
achieved a much better IGD value than GrEA and NSGA-III consumes the highest amount of runtime on all numbers of
on DTLZ6, although these IGD values obtained by KnEA are objectives, which is due to its very intensive computational
still slightly worse than those obtained by HypE and MOEA/D. complexity for repeatedly calculating the HV.
Note, however, that GrEA and NSGA-III outperformed KnEA The runtime of MOEA/D should remain roughly the same
on DTLZ1 and DTLZ3 with more than three objectives. This as the number of objectives increases. The main reason is
may be attributed to the fact that DTLZ1 and DTLZ3 are that MOEA/D decomposes an MOP into a number of single-
multimodal test problems containing a large number of local objective optimization subproblems, where the number of
Pareto optimal fronts and preference over knee points in the subproblems is determined by the predefined population size,
neighborhood easily results in the preference over local Pareto regardless of the number of objectives of the MOP. However,
optimal solutions. This could be partly alleviated by using a from Fig. 6, we can see that the runtime of MOEA/D on
smaller threshold T, which is the predefined maximal ratio of DTLZ test problems increased as the number of objectives
knee points to the nondominated solutions in a nondominated increases, which is attributed to the larger population size on
front. Therefore, for multimodal MOPs, T needs to be chosen problems with an increased number of objectives. The runtime
more carefully to balance exploration and exploitation. More consumed by KnEA, NSGA-III, GrEA, and HypE is expected
detailed discussions on the influence of T on the performance to increase as the number of objectives increases, since GrEA,
of KnEA will be presented in Section IV-D. NSGA-III, and KnEA are all based on nondominated sort-
From the 35 test instances of the DTLZ test suite presented ing and the number of nondominated solutions will increase
in Table VI, we can find that KnEA wins in 11 instances in significantly as the number of objectives increases, while the
terms of IGD, while GrEA wins 5, HypE 5, MOEA/D 7, and computational time for calculating the HV suffers from a
NSGA-III 7. From these results, we can conclude that KnEA dramatic increase when the number of objectives increases.
outperforms HypE, MOEA/D, GrEA, and NSGA-III on DTLZ The rapid increase in runtime of GrEA can be attributed
test problems in terms of IGD, especially for problems with to its environmental selection, where only one solution is
more than three objectives. selected at a time from solutions that cannot be distinguished
Fig. 6 illustrates the runtime of the five algorithms on using dominance comparison, which is quite time-consuming
all DTLZ test problems, where the runtime of an algorithm when the number of nondominated solutions becomes large.
on M objectives is obtained by averaging over the runtimes In KnEA, by contrast, all other nondominated solutions apart

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 772 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO. 6, DECEMBER 2015

TABLE VII
PARAMETER S ETTING FOR T EST P ROBLEMS WFG1 TO WFG9 or close to the best HV was obtained by KnEA, NSGA-III, and
GrEA on all WFG problems with more than three objectives,
especially for WFG5, WFG6, WFG8, and WFG9. On these
four WFG problems, KnEA, NSGA-III, and GrEA obtained a
HV that is at least two times of that obtained by HypE and
MOEA/D. HypE and MOEA/D achieved a good performance
on some WFG test instances with more than three objectives.
Among the five compared algorithms, HypE obtained the best
HV on WFG2 with eight and ten objectives, while MOEA/D
achieved the best HV on WFG3 with four, eight, and ten
objectives.
KnEA performed comparably well with NSGA-III and
from the knee points can be selected at once according to their
GrEA on WFG test problems with more than three objec-
distance to the hyperplane. This saves much time for KnEA
tives, and often better on most WFG test instances when the
compared to GrEA and HypE, particularly when the number
number of objectives is larger than six. For all 18 WFG test
of objectives is large.
instances with eight and ten objectives, KnEA only obtained a
To summarize, we can conclude from Table VI and Fig. 6
slightly worse HV than NSGA-III and GrEA on WFG2 with
that KnEA performs the best among the five compared algo-
eight and ten objectives, WFG3 with eight and ten objectives
rithms. KnEA is computationally also much more efficient
and WFG8 with eight objectives. These results indicate that
than many Pareto-based or performance indicator based popu-
KnEA is more suited to deal with MaOPs with more than
lar MOEAs such as GrEA and HypE, and comparable with
six objectives than GrEA and NSGA-III.
NSGA-III and MOEA/D, which are computationally very
Overall, KnEA performed better than MOEA/D, HypE,
efficient MOEAs.
NSGA-III, and GrEA on the WFG test suite in terms of
HV. KnEA achieved the best HV on 22 test instances out
C. Results on the WFG Suite of 45 WFG test instances considered in this paper, while
The WFG test suite was first introduced in [60] and system- GrEA, NSGA-III, HypE, and MOEA/D achieved the best
atically reviewed and analyzed in [55], which was designed HV on ten instances, three instances, five instances, and
with the aim to introduce a class of difficult benchmark prob- five instances, respectively. Therefore, we can conclude that
lems for evaluating the performance of MOEAs. In this paper, KnEA is very competitive for solving the WFG test func-
we used nine test problems, from WFG1 to WFG9. The param- tions, especially for problems with more than three objectives.
eters of these problems are set as suggested in [55], which are Note that KnEA performed very well for all WFG test func-
listed in Table VII. tions even on the multimodal problems WFG4 and WFG9,
Like in previous work, we compare the quality of the solu- since a small value of T has been adopted in KnEA on
tion sets obtained by the compared algorithms on the nine WFG4 and WFG9 with a large number of objectives, which
WFG test problems in terms of HV, which is another very confirms that a carefully selected small value of T is help-
popular performance indicator that takes both accuracy (close- ful for KnEA to achieve a good performance on multimodal
ness to the true Pareto front) and the diversity of the solution problems.
set into account. Table VIII presents the mean and standard Fig. 7 illustrates the runtime of the five algorithms on all
deviation of the HVs of the five algorithms on WFG1 to WFG test problems, where the runtime of an algorithm on
WFG9, averaging over 20 independent runs, where the best M objectives is obtained by averaging over the runtimes con-
mean among the five algorithms is highlighted. From this sumed by the algorithm for one run on all WFG test problems
table, the following observations can be made. First, MOEA/D, with M objectives. Note that in the figure the runtimes are
HypE, and NSGA-III still achieved a good performance on displayed in logarithm. As can be seen from Fig. 7, we can
WFG test problems with two objectives in terms of HV. find that the average runtime of KnEA is much less than that
MOEA/D and NSGA-III obtained a HV close to the best of GrEA and HypE, comparable to NSGA-III, however, is
one on all WFG test problems with two objectives, while still slightly more than that of MOEA/D. This demonstrates
HypE obtained the best HV on WFG4, WFG5, and WFG9 that the performance of KnEA is very promising in terms of
with two objectives among the five algorithms under compari- runtime.
son. These empirical results confirm that MOEA/D, HypE, and From Table VIII and Fig. 7, we can conclude that over-
NSGA-III are promising algorithms for MOPs with a small all, KnEA showed the most competitive performance on the
number of objectives. For WFG problems with two objectives, WFG test problems. In addition, KnEA is computationally
the performance of GrEA and KnEA is also encouraging, since much more efficient than GrEA and HypE, and comparable
they were able to produce comparable results with those of to NSGA-III and MOEA/D, which is very encouraging.
MOEA/D, HypE, and NSGA-III on all WFG problems with
two objectives.
By contrast, KnEA, NSGA-III, and GrEA performed consis- D. Sensitivity of Parameter T in KnEA
tently much better than MOEA/D and HypE in terms of HV on KnEA has one algorithm specific parameter T, which is
WFG problems with more than three objectives. The best HV used to control the ratio of knee points to the nondominated

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: KNEE POINT-DRIVEN EA FOR MANY-OBJECTIVE OPTIMIZATION 773

TABLE VIII
HV S OF THE F IVE A LGORITHMS ON WFG1 TO WFG9, W HERE THE B EST M EAN FOR E ACH T EST I NSTANCE
IS S HOWN W ITH A G RAY BACKGROUND

solutions in the combined population. In the following, we From the parameter settings in the previous experiments, we
investigate the influence of T on the performance of KnEA, have already noted that T has been set to different values in
which varies from 0.1 to 0.9. Note that 0 < T < 1. KnEA depending on whether the optimization problem has a

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 774 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO. 6, DECEMBER 2015

Fig. 9. IGD values on DTLZ2 of KnEA with different settings for

parameter T, averaging over 20 independent runs.
Fig. 7. Runtime(s) of the five algorithms on all WFG test problems, where
the runtime of an algorithm on M objectives is obtained by averaging over
the runtimes consumed by the algorithm for a run on all M-objective WFG The experimental results on DTLZ2 are summarized in
problems.
Fig. 9, where the mean IGD values for different settings of
parameter T on DTLZ2 with 2, 4, 6, 8, and 10 objectives
averaging over 20 independent runs are presented. We can see
from the figure that the IGD value will first become smaller
as T increases up to 0.6 for the bi-objective DTLZ2 and up to
0.5 for DTLZ2 having more than two objectives. Compared
to the T values that produce the best performance for DTLZ1,
we can conclude that for MOPs that do not have a large num-
ber of local Pareto fronts, T can be set between 0.5 and 0.6,
where a slightly larger T can be used for a smaller number of
objectives.
To summarize the above results, we can conclude that
although the performance of KnEA varies with the value of
Fig. 8. IGD values on DTLZ1 of KnEA with different settings for parameter T, there is a pattern that can be followed to guide
parameter T, averaging over 20 independent runs. the setting for T. For MOPs without a large number of local
Pareto optimal fronts, T can be set to 0.6 for bi-objective prob-
lems, to a value around 0.5 for problems having more than two
objectives. For MOPs with a large number of local Pareto opti-
large number of local Pareto optimal fronts. The main reason
mal fronts, T = 0.5 is recommended for bi- or three-objective
is that a relatively small T is helpful for KnEA to escape from
problems, while for problems with more than three objectives,
local Pareto fronts. For this reason, we consider the setting of
a small value of T is recommended and the larger the number
T on two DTLZ test problems, DTLZ1 and DTLZ2, with the
of objectives is, the smaller the value of T should be used.
former representing a class of optimization problems having a
large number of local Pareto fronts, while the latter represent-
ing a class of test problems that do not have a large number V. C ONCLUSION
of local Pareto optimal fronts. Note that similar results have In this paper, a novel MOEA for solving MaOPs, called
been obtained on other test problems. KnEA, has been proposed. The main idea is to make use of
Fig. 8 shows the results of IGD values for different settings knee points to enhance the search performance of MOEAs
of parameter T on DTLZ1 with 2, 4, 6, 8, and 10 objec- when the number of objectives becomes large. In KnEA, the
tives, averaging over 20 independent runs. Note that in the knee points in the nondominated solutions are preferred to
figure the IGD values are displayed in logarithm. We can see other nondominated solutions in mating selection and envi-
that as T varies from 0.1 to 0.9, the IGD value of KnEA ronmental selection. To the best of our knowledge, this is
on DTLZ1 first decreases, and then will increase again. For the first time that knee points have been used to increase the
DTLZ1 with two and four objectives, the best performance selection pressure in solving MaOPs, thereby improving the
has been achieved when T is around 0.6, while for DTLZ1 convergence performance of Pareto-based MOEAs.
with six objectives, the best performance is achieved when In KnEA, a new adaptive algorithm for identifying knee
T = 0.2, and for DTLZ1 with eight and ten objectives, T = 0.1 points in the nondominated solutions has been developed.
produces the best performance. In general, the experimental While most existing MOEAs for knee points aim to accurately
results confirm that for multimodal MOPs, a relatively small T, locate the knee solutions in the true Pareto front, the proposed
e.g., between 0.1 and 0.4 may be more likely to lead to good adaptive knee point identification algorithm intends to find
performance, particularly when the number of objectives is knee points in the neighborhood of solutions in the nondom-
larger than four. For multimodal MOPs having two to four inated fronts during the optimization, thereby distinguishing
objectives, T can be set to between 0.5 and 0.6. some of the nondominated solutions from others. To this end,

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: KNEE POINT-DRIVEN EA FOR MANY-OBJECTIVE OPTIMIZATION 775

the adaptive strategy attempts to maintain a proper ratio of the [10] D. W. Corne, N. R. Jerram, J. D. Knowles, and M. J. Oates, “PESA-II:
identified knee points to all nondominated solutions in each Region-based selection in evolutionary multi-objective optimization,”
in Proc. 2001 Genet. Evol. Comput. Conf., San Francisco, CA, USA,
front by adjusting the size of the neighborhood of each solu- pp. 283–290.
tion in which the solution having the maximum distance to [11] J. D. Knowles and D. W. Corne, “M-PAES: A memetic algorithm
the hyperplane is identified as the knee point. In this way, the for multiobjective optimization,” in Proc. IEEE Congr. Evol. Comput.,
La Jolla, CA, USA, 2000, pp. 325–332.
preference over knee points in selection will not only acceler- [12] K. Narukawa and T. Rodemann, “Examining the performance of evolu-
ate the convergence performance but also the diversity of the tionary many-objective optimization algorithms on a real-world applica-
population. tion,” in Proc. 6th Int. Conf. Genet. Evol. Comput., Kitakyushu, Japan,
2012, pp. 316–319.
Comparative experimental results with four popular [13] R. J. Lygoe, M. Cary, and P. J. Fleming, “A real-world application of
MOEAs, namely, MOEA/D, HypE, GrEA, and NSGA-III a many-objective optimisation complexity reduction process,” in Proc.
demonstrate that the proposed KnEA significantly outper- 7th Int. Conf. Evol. Multi-Criterion Optim., Sheffield, U.K., 2013,
pp. 641–655.
forms MOEA/D and HypE, and is comparable with GrEA [14] K. Ikeda, H. Kita, and S. Kobayashi, “Failure of Pareto-based MOEAs:
and NSGA-III on MaOPs with more than three objectives. Does non-dominated really mean near to optimal?” in Proc. IEEE Congr.
Most encouragingly, KnEA is computationally much more Evol. Comput., Seoul, Korea, 2001, pp. 957–962.
[15] D. Brockhoff et al., “On the effects of adding objectives to plateau
efficient compared with other Pareto-based MOEAs such as functions,” IEEE Trans. Evol. Comput., vol. 13, no. 3, pp. 591–603,
GrEA and performance indicator based MOEAs such as HypE. Jun. 2009.
Therefore, the overall performance of KnEA is highly com- [16] O. Schütze, A. Lara, and C. A. C. Coello, “On the influence of the
number of objectives on the hardness of a multiobjective optimization
petitive compared to the state-of-the-art MOEAs for solving problem,” IEEE Trans. Evol. Comput., vol. 15, no. 4, pp. 444–455,
MaOPs. Aug. 2011.
This paper demonstrates that the idea of using knee points to [17] L. S. Batista, F. Campelo, F. G. Guimaraes, and J. A. Ramírez, “A com-
parison of dominance criteria in many-objective optimization problems,”
increase the selection pressure for MaOPs is very promising. in Proc. IEEE Congr. Evol. Comput., New Orleans, LA, USA, 2011,
Further work on developing more effective and computation- pp. 2359–2366.
ally more efficient algorithms for identifying knee solutions is [18] V. Khare, X. Yao, and K. Deb, “Performance scaling of multiobjective
evolutionary algorithms,” in Proc. 2nd Int. Conf. Evol. Multi-Criterion
highly desirable. In KnEA, nondominated solutions other than Optim., Faro, Portugal, 2003, pp. 376–390.
the knee points have been selected according to their distance [19] D. W. Corne and J. D. Knowles, “Techniques for highly multiobjective
to the hyperplane. This idea has been shown to be effective in optimisation: Some nondominated points are better than others,” in Proc.
9th Conf. Genet. Evol. Comput., London, U.K., 2007, pp. 773–780.
KnEA, however, the performance of KnEA could be further [20] H. Ishibuchi, N. Tsukamoto, and Y. Nojima, “Evolutionary many-
improved by introducing criteria other than the distance to the objective optimization: A short review,” in Proc. IEEE Congr. Evol.
hyperplane. Finally, the performance of KnEA remains to be Comput., Hong Kong, 2008, pp. 2419–2426.
[21] Z. He and G. G. Yen, “Ranking many-objective evolutionary algo-
verified on real-world MaOPs. rithms using performance metrics ensemble,” in Proc. IEEE Congr. Evol.
Comput., Cancun, Mexico, 2013, pp. 2480–2487.
[22] M. Laumanns, L. Thiele, K. Deb, and E. Zitzler, “Combining con-
vergence and diversity in evolutionary multiobjective optimization,”
R EFERENCES Evol. Comput., vol. 10, no. 3, pp. 263–282, 2002.
[1] J. G. Herrero, A. Berlanga, and J. M. M. López, “Effective evolutionary [23] D. Hadka and P. Reed, “Borg: An auto-adaptive many-objective
algorithms for many-specifications attainment: Application to air traf- evolutionary computing framework,” Evol. Comput., vol. 21, no. 2,
fic control tracking filters,” IEEE Trans. Evol. Comput., vol. 13, no. 1, pp. 231–259, 2013.
pp. 151–168, Feb. 2009. [24] X. Zou, Y. Chen, M. Liu, and L. Kang, “A new evolutionary algorithm
for solving many-objective optimization problems,” IEEE Trans. Syst.,
[2] H. Ishibuchi and T. Murata, “A multi-objective genetic local search
algorithm and its application to flowshop scheduling,” IEEE Trans. Man, Cybern. B, Cybern., vol. 38, no. 5, pp. 1402–1412, Oct. 2008.
Syst., Man, Cybern. C, Appl. Rev., vol. 28, no. 3, pp. 392–403, [25] G. Wang and H. Jiang, “Fuzzy-dominance and its application in evolu-
tionary many objective optimization,” in Proc. Int. Conf. Comput. Intell.
Aug. 1998.
Security Workshops, Harbin, China, 2007, pp. 195–198.
[3] S. H. Yeung, K. F. Man, K. M. Luk, and C. H. Chan, “A trapeizform
[26] F. di Pierro, S.-T. Khu, and D. A. Savic, “An investigation on prefer-
U-slot folded patch feed antenna design optimized with jumping genes
ence order ranking scheme for multiobjective evolutionary optimization,”
evolutionary algorithm,” IEEE Trans. Antennas Propag., vol. 56, no. 2,
IEEE Trans. Evol. Comput., vol. 11, no. 1, pp. 17–45, Feb. 2007.
pp. 571–577, Feb. 2008.
[27] M. Köppen and K. Yoshida, “Substitute distance assignments in
[4] J. Handl, D. B. Kell, and J. Knowles, “Multiobjective optimization in NSGA-II for handling many-objective optimization problems,” in Proc.
bioinformatics and computational biology,” IEEE/ACM Trans. Comput. 4th Int. Conf. Evol. Multi-Criterion Optim., Matsushima, Japan, 2007,
Biol. Bioinformat., vol. 4, no. 2, pp. 279–292, Apr. 2007. pp. 727–741.
[5] A. Ponsich, A. L. Jaimes, and C. A. C. Coello, “A survey on multiobjec- [28] A. López, C. A. C. Coello, A. Oyama, and K. Fujii, “An alternative
tive evolutionary algorithms for the solution of the portfolio optimization preference relation to deal with many-objective optimization problems,”
problem and other finance and economics applications,” IEEE Trans. in Proc. 7th Int. Conf. Evol. Multi-Criterion Optim., Sheffield, U.K.,
Evol. Comput., vol. 17, no. 3, pp. 321–344, Jun. 2013. 2013, pp. 291–306.
[6] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and eli- [29] S. Yang, M. Li, X. Liu, and J. Zheng, “A grid-based evolutionary algo-
tist multi-objective genetic algorithm: NSGA-II,” IEEE Trans. Evol. rithm for many-objective optimization,” IEEE Trans. Evol. Comput.,
Comput., vol. 6, no. 2, pp. 182–197, Apr. 2002. vol. 17, no. 5, pp. 721–736, Oct. 2013.
[7] E. Zitzler, M. Laumanns, and L. Thiele, “SPEA2: Improving the strength [30] N. Beume, B. Naujoks, and M. Emmerich, “SMS-EMOA: Multiobjective
Pareto evolutionary algorithm for multiobjective optimization,” in Proc. selection based on dominated hypervolume,” Eur. J. Oper. Res., vol. 181,
5th Conf. Evol. Methods Design Optim. Control Appl. Ind. Probl., no. 3, pp. 1653–1669, 2007.
Athens, Greece, 2001, pp. 95–100. [31] J. Bader and E. Zitzler, “HypE: An algorithm for fast hypervolume-based
[8] E. Zitzler and S. Künzli, “Indicator-based selection in multiobjective many-objective optimization,” Evol. Comput., vol. 19, no. 1, pp. 45–76,
search,” in Proc. 8th Int. Conf. Parallel Probl. Solving Nat., Birmingham, 2011.
U.K., 2004, pp. 832–842. [32] D. K. Saxena and K. Deb, “Dimensionality reduction of objectives and
[9] Q. Zhang and H. Li, “MOEA/D: A multiobjective evolutionary algorithm constraints in multi-objective optimization problems: A system design
based on decomposition,” IEEE Trans. Evol. Comput., vol. 11, no. 6, perspective,” in Proc. IEEE Congr. Evol. Comput., Hong Kong, 2008,
pp. 712–731, Dec. 2007. pp. 3204–3211.

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.
 776 IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 19, NO. 6, DECEMBER 2015

[33] H. K. Singh, A. Isaacs, and T. Ray, “A Pareto corner search evolutionary [56] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms.
algorithm and dimensionality reduction in many-objective optimization New York, NY, USA: Wiley, 2001.
problems,” IEEE Trans. Evol. Comput., vol. 15, no. 4, pp. 539–556, [57] Q. Zhang, A. Zhou, and Y. Jin, “RM-MEDA: A regularity model-based
Aug. 2011. multiobjective estimation of distribution algorithm,” IEEE Trans. Evol.
[34] K. Deb, A. Sinha, P. J. Korhonen, and J. Wallenius, “An interactive evo- Comput., vol. 12, no. 1, pp. 41–63, Feb. 2008.
lutionary multi-objective optimization method based on progressively [58] M. Li and J. Zheng, “Spread assessment for evolutionary multiobjec-
approximated value functions,” IEEE Trans. Evol. Comput., vol. 14, tive optimization,” in Proc. 5th Int. Conf. Evol. Multi-Criterion Optim.,
no. 5, pp. 723–739, Oct. 2010. Nantes, France, 2009, pp. 216–230.
[35] R. Wang, R. C. Purshouse, and P. J. Fleming, “Preference-inspired [59] L. M. S. Russo and A. P. Francisco, “Quick hypervolume,” IEEE Trans.
coevolutionary algorithms for many-objective optimization,” IEEE Evol. Comput., vol. 18, no. 4, pp. 481–502, Aug. 2014.
Trans. Evol. Comput., vol. 17, no. 4, pp. 474–494, Aug. 2013. [60] S. Huband, L. Barone, L. While, and P. Hingston, “A scalable multiob-
[36] K. Deb and H. Jain, “An evolutionary many-objective optimization jective test problem toolkit,” in Proc. 3th Int. Conf. Evol. Multi-Criterion
algorithm using reference-point-based nondominated sorting approach, Optim., Guanajuato, Mexico, 2005, pp. 280–294.
part I: Solving problems with box constraints,” IEEE Trans. Evol.
Comput., vol. 18, no. 4, pp. 577–601, Aug. 2014.
[37] H. Ishibuchi, N. Akedo, and Y. Nojima, “Relation between neighborhood
size and MOEA/D performance on many-objective problems,” in Proc.
Evol. Multi-Criterion Optim., Sheffield, U.K., 2013, pp. 459–474. Xingyi Zhang received the B.Sc. degree from
[38] L. B. Said, S. Bechikh, and K. Ghédira, “The r-dominance: A new Fuyang Normal College, Anhui, China, in 2003,
dominance relation for interactive evolutionary multicriteria decision and the M.Sc. and Ph.D. degrees from Huazhong
making,” IEEE Trans. Evol. Comput., vol. 14, no. 5, pp. 801–818, University of Science and Technology, Wuhan,
Oct. 2010. China, in 2006 and 2009, respectively.
[39] J. Molina, L. V. Santana, A. G. Hernández-Díaz, C. A. C. Coello, He is an Associate Professor with the
and R. Caballero, “g-dominance: Reference point based dominance School of Computer Science and Technology,
for multiobjective metaheuristics,” Eur. J. Oper. Res., vol. 197, no. 2, Anhui University, Anhui. His research interests
pp. 685–692, 2009. include unconventional models and algorithms
[40] L. Thiele, K. Miettinen, P. J. Korhonen, and J. Molina, “A of computation, multiobjective optimization, and
preference-based evolutionary algorithm for multi-objective optimiza- membrane computing.
tion,” Evol. Comput., vol. 17, no. 3, pp. 411–436, 2009.
[41] I. Das, “On characterizing the ‘knee’ of the Pareto curve based
on normal-boundary intersection,” Struct. Optim., vol. 18, nos. 2–3,
pp. 107–115, 1999.
[42] J. Branke, K. Deb, H. Dierolf, and M. Osswald, “Finding knees in multi- Ye Tian received the B.Sc. degree from Anhui
objective optimization,” in Proc. 8th Int. Conf. Parallel Probl. Solving University, Anhui, China, in 2012, where he is cur-
Nat., Birmingham, U.K., 2004, pp. 722–731. rently working toward the master’s degree from the
[43] L. Rachmawati and D. Srinivasan, “A multi-objective genetic algorithm School of Computer Science and Technology.
with controllable convergence on knee regions,” in Proc. IEEE Congr. His research interests include multiobjective opti-
Evol. Comput., Vancouver, BC, Canada, 2006, pp. 1916–1923. mization methods and their application.
[44] L. Rachmawati and D. Srinivasan, “A multi-objective evolutionary algo-
rithm with weighted-sum niching for convergence on knee regions,”
in Proc. 8th Int. Conf. Genet. Evol. Comput., Seattle, WA, USA, 2006,
pp. 749–750.
[45] O. Schütze, M. Laumanns, and C. A. C. Coello, “Approximating the
knee of an MOP with stochastic search algorithms,” in Proc. 10th
Int. Conf. Parallel Probl. Solving Nat., Dortmund, Germany, 2008,
pp. 795–804.
Yaochu Jin (M’98–SM’02) received the B.Sc.,
[46] S. Bechikh, L. B. Said, and K. Ghédira, “Searching for knee regions M.Sc., and Ph.D. degrees from Zhejiang University,
in multi-objective optimization using mobile reference points,” in Proc. Hangzhou, China, in 1988, 1991, and 1996, respec-
ACM Symp. Appl. Comput., Sierre, Switzerland, 2010, pp. 1118–1125. tively, and the Dr.-Ing. degree from Ruhr University,
[47] K. Deb, J. Sundar, U. Bhaskara, and S. Chaudhuri, “Reference point Bochum, Germany, in 2001.
based multi-objective optimization using evolutionary algorithms,” Int. He is a Professor of Computational Intelligence,
J. Comput. Intell. Res., vol. 2, no. 3, pp. 635–642, 2006. Department of Computing, University of Surrey,
[48] K. Deb and S. Gupta, “Understanding knee points in bicriteria problems Guildford, U.K., where he heads the Nature Inspired
and their implications as preferred solution principles,” Eng. Optim., Computing and Engineering Group. He is also
vol. 43, no. 11, pp. 1175–1204, 2011. a Finland Distinguished Professor funded by the
[49] T. Tušar and B. Filipič, “Scaling and visualizing multiobjective opti- Finnish Agency for Innovation (Tekes). His research
mization test problems with knees,” in Proc. 15th Int. Multiconf. Inf. interests include the interdisciplinary areas that bridge the gap between
Soc., Ljubljana, Slovenia, 2012, pp. 155–158. computational intelligence, computational neuroscience, computational sys-
[50] L. While, P. Hingston, L. Barone, and S. Huband, “A faster algorithm tems biology, and nature-inspired, real-world-driven problem-solving. He has
for calculating hypervolume,” IEEE Trans. Evol. Comput., vol. 10, no. 1, (co)edited five books and three conference proceedings, authored a mono-
pp. 29–38, Feb. 2006. graph, and (co)authored over 200 peer-reviewed journal and conference
[51] X. Zhang, Y. Tian, R. Cheng, and Y. Jin, “An efficient approach papers. He holds eight U.S., EU, and Japan patents. His research is funded
to nondominated sorting for evolutionary multiobjective optimization,” by EC FP7, U.K. Engineering and Physical Sciences Research Council,
IEEE Trans. Evol. Comput., vol. 19, no. 2, pp. 201–213, 2015. and industry. He has delivered 20 invited keynote speeches at international
[52] K. Deb and R. B. Agrawal, “Simulated binary crossover for continuous conferences.
search space,” Compl. Syst., vol. 9, no. 4, pp. 115–148, 1995. Dr. Jin received the Best Paper Award of the 2010 IEEE Symposium
[53] K. Deb and M. Goyal, “A combined genetic adaptive search (GeneAS) on Computational Intelligence in Bioinformatics and Computational
for engineering design,” Comput. Sci. Informat., vol. 26, no. 4, Biology and the 2014 IEEE Computational Intelligence Magazine
pp. 30–45, 1996. Outstanding Paper Award. He is an Associate Editor or an Editorial
[54] K. Deb, L. Thiele, M. Laumanns, and E. Zitzler, “Scalable test prob- Board Member of IEEE T RANSACTIONS ON C YBERNETICS, IEEE
lems for evolutionary multi-objective optimization,” Dept. Comput. Eng. T RANSACTIONS ON NANOBIOSCIENCE, IEEE Computational Intelligence
Netw. Lab., ETH Zurich, Zurich, Switzerland, TIK-Tech. Rep. 112, Magazine, Evolutionary Computation, BioSystems, International Journal of
2001. Fuzzy Systems, Soft Computing, Memetic Computing, and Natural Computing.
[55] S. Huband, P. Hingston, L. Barone, and L. While, “A review of multiob- He is an IEEE Distinguished Lecturer and a Vice President for Technical
jective test problems and a scalable test problem toolkit,” IEEE Trans. Activities of the IEEE Computational Intelligence Society. He is a fellow of
Evol. Comput., vol. 10, no. 5, pp. 477–506, Oct. 2006. the British Computer Society.

Authorized licensed use limited to: K. Ramakrishnan Health and Educational Trust. Downloaded on July 19,2024 at 09:55:20 UTC from IEEE Xplore. Restrictions apply.